Académique Documents
Professionnel Documents
Culture Documents
We would like to thank Richard Curtin and Guhan Venkatu for help with data sources, and
Simon Gilchrist, Robert King, and John Williams for their comments. Doug Geyser and
Cameron Shelton provided research assistance. Ricardo Reis is grateful to the Fundacao
Ciencia e Tecnologia, Praxis XXI, for financial support.
2. Inflation Expectations
Most macroeconomic models argue that inflation expectations are a crucial factor in the inflation process. Yet the nature of these expectationsin
the sense of precisely stating whose expectations, over which prices, and
over what horizonis not always discussed with precision. These are crucial issues for measurement.
The expectations of wage- and price-setters are probably the most relevant. Yet it is not clear just who these people are. As such, we analyze data
from three sources. The Michigan Survey of Consumer Attitudes and
Behavior surveys a cross section of the population about their expectations
over the next year. The Livingston Survey and the Survey of Professional
Forecasters (SPF) covers more sophisticated analystseconomists working
Michigan survey
Livingston survey
Survey
population
Cross section of
the general
public
Academic,
business, finance,
market, and
labor economists
Market
economists
Survey
organization
Survey Research
Center, University
of Michigan
Originally Joseph
Livingston, an
economic journalist;
currently the
Philadelphia Fed
Originally
ASA/NBER;
currently the
Philadelphia Fed
Average
number of
respondents
Roughly 1000-3000
per quarter to 1977,
then 500-700 per
month to present
48 per survey
(varies from
14-63)
34 per survey
(varies from 9-83)
Starting date
Qualitative
questions: 1946
Ql1; quantitative
responses: January
1978
Periodicity
Semi-annual
Quarterly
Inflation
expectation
Expected change
in prices over the
next 12 months
Consumer Price
Index (this quarter,
in 2 quarters, in
4 quarters)
Q4;
CPI inflation:
1981, Q3
Michigan Experimental
Livingston
15-1
10-
0-
15-
1950
1960
1970
1980
1990
2000 1950
1960
1970
1980
1990
2000
Expected Inflation
rather than at the time the forecast was made. Two striking features
emerge from these plots. First, each series yields relatively accurate inflation forecasts. And second, despite the different populations being surveyed, they all tell a somewhat similar story.
By simple measures of forecast accuracy, all three surveys appear to be
quite useful. Table 2 shows two common measures of forecast accuracy:
the square root of the average squared error (RMSE) and the mean
absolute error (MAE). In each case we report the accuracy of the median
expectation in each survey, both over their maximal samples and for a
common sample (September 1982-March 2002).
Panel A of the table suggests that inflation expectations are relatively
accurate. As the group making the forecast becomes increasingly sophisticated, forecast accuracy appears to improve. However, Panel B suggests that
these differences across groups largely reflect the different periods over
which each survey has been conducted. For the common sample that all five
measures have been available, they are all approximately equally accurate.
Of course, these results reflect the fact that these surveys have a similar
central tendency, and this fact reveals as much as it hides. Figure 2 presents simple histograms of expected inflation for the coming year as of
December 2002.
Here, the differences among these populations become starker. The left
panel pools responses from the two surveys of economists and shows
some agreement on expectations, with most respondents expecting inflation in the VA to 3% range. The survey of consumers reveals substantially
greater disagreement. The interquartile range of consumer expectations
stretches from 0 to 5%, and this distribution shows quite long tails, with
5% of the population expecting deflation, while 10% expect inflation of at
Table 2 INFLATION FORECAST ERRORS
Michigan
Michigan
experimental
Livingston
SPF-GDP
deflator
SPF-CPI
1954, Q42002, Q l
2.32%
1.77%
1954, H l 2001, H2
1.99%
1.38%
1969, Q42002, Q l
1.62%
1.22%
1982, Q 3 2002, Q l
1.29%
0.97%
1.10%
0.91%
1.29%
0.97%
1.07%
0.85%
1.24%
0.95%
1.28%
0.97%
Consumers
Michigan Survey
Empirical Distribution
Empirical Distribution
10-
10-
0- 1 0
1
2
3
4
Expected Inflation over the Year to December 2003, %
-5.0
-2.5
0.0
2.5
5.0
7.5
10.0
Expected Inflation over the Year to December 2003, %
Expectations < -5% and > 10% truncated lo endpoints.
f 4H
a
Michigan Survey
2H
1950
1960
1970
1980
1990
2000
Year
2.5H
S)2.0*15H
1.0^0
o.oH
1950
1970
1980
Livingston Survey
1990
2000
1990
2000
Year
Inflation Rate
2
5i
o1950
1960
1970
1980
Year
Livingston
0.682
0.809
1.000
0.391
1.000
0.700
0.667
0.502
0.231
0.712
0.702
Michigan
SPF-GDP
deflator
SPF-CPI
1.000
0.688
1.000
1.000
0.865
1.000
Michigan
Michigan
experimental
Livingston
SPF-GDP
deflator
SPF-CPI
1.000
Michigan
Michigan
experimental
Livingston
SPF-GDP
deflator
SPF-CPI
1.000
0.729
0.869
1.000
0.813
1.000
0.850
0.868
0.690
0.308
0.889
0.886
1. Underlying data are quarterly. They are created by taking averages of monthly Michigan data and by linearly interpolating half-yearly Livingston data.
A final source of data on disagreement comes from the range of forecasts within the Federal Open Market Committee (FOMC), as published
biannually since 1979 in the Humphrey-Hawkins testimony.3 Individuallevel data are not released, so we simply look to describe the broad pattern of disagreement among these experts. Figure 4 shows a rough (and
statistically significant) correspondence between disagreement among
policymakers and disagreement among professional economists. The correlation of the range of FOMC forecasts with the interquartile range of the
Livingston population is 0.34, 0.54 or 0.63, depending on which of the
three available FOMC forecasts we use. While disagreement among Fedwatchers rose during the Volcker disinflation, the range of inflation forecasts within the Fed remained largely constantthe correlation between
disagreement among FOMC members and disagreement among professional forecasters is substantially higher after 1982.
We believe that we have now established three important patterns in the
data. First, there is substantial disagreement within both naive and expert
3. We are grateful to Simon Gilchrist for suggesting this analysis to us. Data were drawn
from Gavin (2003) and updated using recent testimony published at http://www.federalreserve.gov/boarddocs/ hh/ (accessed December 2003).
populations about the expected future path of inflation. Second, there are
larger levels of disagreement among consumers than exists among experts.
And third, even though professional forecasters, economists, and the general population show different degrees of disagreement, this disagreement
tends to exhibit similar time-series patterns, albeit of a different amplitude.
One would therefore expect to find that the underlying causes behind this
disagreement are similar across all three datasets.
91
NM)O
so
1
2
3
IQR of Livingston Forecasts (%)
83
' *}
'3-
88A84
A 93
2u
M M 84 80
194 B85
MB87
BS1
82
2
u
A9A82
80
A 0190
AA900
A 95
A96
A 97
Correlation (Whole Sample) = 0.34
Correlation (Post-1982) = 0.74
17)1
mm
60
jjo0
o0
1
2
IQR of Livingston Forecasts (%)
Humphrey-Hawkins testimony in February and July provides forecasts for inflation over the calendar year.
Inflation concept varies.
a: mean error
(Constant only)
Michiganexperimental
Livingston
SPF (GDP
deflator)
-0.09%
(0.34)
0.63%**
(0.30)
-0.02%
(0.29)
=a
0.42%
(0.29)
0.349**
(.161)
-1.016%*
(.534)
0.197
Yes
(p = 0.088)
-0.060
(.207)
-0.182%
(.721)
-0.003
No
(p = 0.956)
0.011
(.142)
0.595%
(.371)
-0.011
Yes
(p = 0.028)
a: constant
Adj. R2
0.371**
(0.158)
0.096%
(0.183)
0.164
a: constant
-0.816%
(0.975)
0.801***
P : Ef-12 [nt]
(0.257)
y: inflation(_13
-0.218*
(0.121)
K: Treasury billt_13
-0.165**
(0.085)
8: u n e m p l o y m e n t ^ 0.017
(0.126)
Reject eff.? y = K = 8 = 0 Yes
(p-value)
(p = 0.049)
Adjusted R2
0.293
Sample
Periodicity
N
+ P (7if_12 - Et_24'i^-12)
.580***
(0.115)
0.005%
(0.239)
0.334
0.026
(.128)
-0.132%
(.530)
-0.007
No
(p = 0.969)
0.490***
(0.132)
0.302%
(0.210)
0.231
-
Et_12%t
0.640***
(0.224)
-0.032%
(0.223)
0.375
= a
P (_12 [7tJ
0.242%
(1.143)
-0.554***
(0.165)
0.610***
(0.106)
-0.024
(0.102)
-0.063
(0.156)
Yes
(p = 0.000)
0.382
4.424%***
(0.985)
0.295
(0.283)
0.205
(0.145)
-0.319***
(0.106)
-0.675***
(0.175)
Yes
(p = 0.000)
0.306
3.566%***
(0.970)
0.287
(0.308)
0.200
(0.190)
-0.321***
(0.079)
-0.593***
(0.150)
Yes
(p = 0.000)
0.407
1954, Q 4 2002, Q l
Quarterly
169
1954, H l 2001, H2
Semiannual
96
1969, Q42002, Ql
Quarterly
125
1. ***, ** and * denote statistical significance at the 1%, 5%, and 10% levels, respectively (Newey-West standard errors in parentheses; correcting for autocorrelation up to one year).
these results probably capture the general flavor of the existing empirical
literature, if not the somewhat stronger arguments made by individual
authors. Bias exists but is typically small. Forecasts are typically inefficient, though not in all surveys: while the forecast errors of economists are
not predictable based merely on their forecasts, those of consumers are.
All four data series show substantial evidence that forecast errors made a
year ago continue to repeat themselves, and that recent macroeconomic
data is not adequately reflected in inflation expectations.
We now turn to analyzing whether the data are consistent with adaptive expectations, probably the most popular alternative to rational expectations in the literature. The simplest backward-looking rule invokes the
prediction that expected inflation over the next year will be equal to inflation over the past year. Ball (2000) suggests a stronger version, whereby
agents form statistically optimal univariate inflation forecasts. The test in
Table 5 is a little less structured, simply regressing median inflation expectations against the last eight nonoverlapping, three-month-ended inflation observations. We add the unemployment rate and short-term interest
rates to this regression, finding that these macroeconomic aggregates also
help predict inflation expectations. In particular, it is clear that when the
unemployment rate rises over the quarter, inflation expectations fall further than adaptive expectations might suggest. This suggests that consumers employ a more sophisticated model of the economy than assumed
in the simple adaptive expectations model.
Consequently we are left with a somewhat negative resultobserved
inflation expectations are consistent with neither the sophistication of
rational expectations nor the naivete of adaptive expectations. This finding holds for our four datasets, and it offers a reasonable interpretation of
the prior literature on inflation expectations. The common thread to these
results is that inflation expectations reflect partial and incomplete updating in response to macroeconomic news. We shall argue in Section 5 that
these results are consistent with models in which expectations are not
updated at every instant, but rather in which updating occurs in a staggered fashion. A key implication is that disagreement will vary with
macroeconomic conditions.
00 LO
LO CN CO
T-H
ON IN
LO T1
ON ON
CO
o o
dodo
cr
o oo
co -^
U
w
ft
X
w
o
LO
J
LO CN LO LO
LT) ON LO 00
t N T-H O 1<
d d r-t d
I
'
'
* ^
_LO
CO 00 O
ON
^ LO o ^
rH O i-J O
f^ f~^i f^i
rn
a;
.. >5
CN
f^i
ft
LO LO
CO 00
o
d d
00 LO
CO 00 LO
CN
LO
CO CN CN CN
LO CO LO CN
O T< O i;
,H
0 dd d
LO
31
a-
ft*
>
a
jS
-a
c
u
w
*
*
CO
o
o
d
ft
X
*
*
CO
CO
LO
00 CO
CN LO CN
T1
o.d d
.94
I
* LfT
ON
*
o
11
o p TON-H
d
d d II
1
"' -
LO 00
CO CO
w
w
>
^H
(^
ft
CN
CN
ON
CN
H
ft
tin
In 3
OH
en
H
o
B
LO
CO
X,
w
o
CO
OH
H->
ion
<X)
13
Int lation
9.0-
6.0-
6.0-
3.0-
3.0-
nge of Expect
o.o-
o.o-5
10
15
-5
Economists: Livingston
3.0-
10
15
Economists :SPF
3.0-
53
80
2.0-
70
73
78
2.0-
74
53
75
75 >yl>^*'^
78 74y
79
79
1.0-
7(P07
8*8fi^^Ma
?73
l
_
5
10
15
-5
)^* 1 90
9-
9-
6-
6-
3-
3-
Economists^-Livingston
3-
73 74
Economists: SPF
3-
a* 2
275 80
1-
-5
IO.O-
-4
-2
2.0-.4
Michigan (smoothed)
2.5-
A.
2-
1.501950
% 4.0-
a.
SPF (smoothed)
1960
1970
1980
Year
1990
2000
20-
10-
5-
o-
1970
1980
Year
1990
2000
7? 8%)
82
9-
^ ?TV
81 82
^ A*
73
73
7
71 7^ 9 9
74
9
V ^
>*^
6-
a,
x
Economists: Livingston
Economists: SPF
80
Pi
74
2-
Jr
78
787g ^r
81 8i 7 5 7 * 0 / ^
00
10
15
-5
10
15
Michiganexperimental
Livingston
SPF (GDP
deflator)
Inflation rate
AInflation-squared
Output gap
Relative price
variability
0.441***
(0.028)
18.227***
(2.920)
0.176
(0.237)
0.665***
(0.056)
0.228***
(0.036)
1.259**
(0.616)
-0.047
(0.092)
0.473***
(0.091)
0.083***
(0.016)
2.682***
(0.429)
0.070**
(0.035)
0.117**
(0.046)
0.092***
(0.013)
2.292**
(0.084)
-0.001
(0.029)
0.132
(0.016)
Panel B: regressions controlling for the inflation rate ([each cell represents a separate
regression)
AInflation-squared
Output gap
Relative price
variability
10.401***
(1.622)
0.415***
(0.088)
0.268***
(0.092)
0.814
(0.607)
0.026
(0.086)
0.210
(0.135)
2.051***
(0.483)
-0.062**
(0.027)
0.085**
(0.042)
-0.406
(0.641)
-0.009
(0.013)
0.099***
(0.020)
0.066***
(0.013)
1.663**
(0.737)
0.020
(0.032)
0.095***
(0.015)
-0.305
(0.676)
-0.007
(0.014)
Inflation rate
AInflation-squared
Output gap
0.408***
(0.028)
7.062***
(1.364)
0.293***
(0.066)
0.217***
(0.034)
0.789
(0.598)
0.017
(0.079)
Inflation rate
AInflation-squared
Output gap
Relative price
variability
0.328***
(0.034)
5.558***
(1.309)
0.336***
(0.067)
0.237***
(0.079)
0.204***
(0.074)
-0.320
(2.431)
-0.061
(0.117)
0.210
(0.159)
0.044**
(0.018)
1.398
(0.949)
0.013
(0.039)
0.062
(0.038)
0.037***
(0.011)
-0.411
(0.624)
0.006
(0.018)
0.100***
(0.022)
1. *** and ** denote statistical significance at the 1% and 5% levels, respectively (Newey-West standard
errors in parentheses; correcting for autocorrelation up to one year).
in inflation is highly correlated with disagreement in bivariate regressions, and controlling for the inflation rate and other macroeconomic variables only slightly weakens this effect. Adding the relative price
variability term further weakens this effect. Relative price variability is a
consistently strong predictor of disagreement across all specifications.
These results are generally stronger for the actual Michigan data than for
the experimental series, and they are generally stronger for the Livingston
series than for the SPF. We suspect that both facts reflect the relative role
of measurement error. Finally, while the output gap appears to be related
to disagreement in certain series, this finding is not robust either across
data series or to the inclusion of controls.
In sum, our analysis of the disagreement data has estimated that disagreement about the future path of inflation tends to:
Rise with inflation.
Rise when inflation changes sharplyin either direction.
Rise in concert with dispersion in rates of inflation across commodity
groups.
Show no clear relationship with measures of real activity.
Finally, we end this section with a note of caution. None of these findings
necessarily reflect causality and, in any case, we have deliberately been
quite loose in even speaking about the direction of likely causation.
However, we believe that these findings present a useful set of stylized
facts that a theory of macroeconomic dynamics should aim to explain.
5. Theories of Disagreement
Most theories in macroeconomics have no disagreement among agents. It
is assumed that everyone shares the same information and that all are
endowed with the same information-processing technology. Consequently, everyone ends up with the same expectations.
A famous exception is the islands model of Robert Lucas (1973).
Producers are assumed to live in separate islands and to specialize in producing a single good. The relative price for each good differs by islandspecific shocks. At a given point in time, producers can observe the price
only on their given islands and from it, they must infer how much of it is
idiosyncratic to their product and how much reflects the general price level
that is common to all islands. Because agents have different information,
they have different forecasts of prices and hence inflation. Since all will
inevitably make forecast errors, unanticipated monetary policy affects real
output: following a change in the money supply, producers attribute some
tion, regardless of how long it has been since the last update. The VAR is
then used to produce estimates of future annual inflation in the United
States given information at different points in the past. To each of these
forecasts, we attribute a frequency as dictated by the process just
described. This generates at each point in time a full cross-sectional distribution of annual inflation expectations. We use the predictions from
1954 onward, discarding the first few years in the sample when there are
not enough past observations to produce nondegenerate distributions.
We compare the predicted distribution of inflation expectations by the
sticky-information model to the distribution we observe in the survey
data. To do so meaningfully, we need a relatively long sample period. This
leads us to focus on the Livingston and the Michigan experimental series,
which are available for the entire postwar period.
The parameter governing the rate of information updating in the economy, X, is chosen to maximize the correlation between the interquartile
range of inflation expectations in the survey data with that predicted by
the model. For the Livingston Survey, the optimal X is 0.10, implying that
the professional economists surveyed are updating their expectations
about every 10 months, on average. For the Michigan series, the value of
X that maximizes the correlation between predicted and actual dispersion
is 0.08, implying that the general public updates their expectations on
average every 12.5 months. These estimates are in line with those
obtained by Mankiw and Reis (2003), Carroll (2003a), and Khan and Zhu
(2002). These authors employ different identification schemes and estimate that agents update their information sets once a year, on average.
Our estimates are also consistent with the reasonable expectation that
people in the general public update their information less frequently than
professional economists do. It is more surprising that the difference
between the two is so small.
A first test of the model is to see to what extent it can predict the dispersion in expectations over time. Figure 10 plots the evolution of the
interquartile range predicted by the sticky-information model, given the
history of macroeconomic shocks and VAR-type updating, and setting
X = 0.1. The predicted interquartile range matches the key features of the
Livingston data closely, and the two series appear to move closely
together. The correlation between them is 0.66. The model is also successful at matching the absolute level of disagreement. While it overpredicts
dispersion, it does so only by 0.18 percentage points on average.
The sticky-information model also predicts the time-series movement in
disagreement among consumers. The correlation between the predicted
and actual series is 0.80 for the actual Michigan data and 0.40 for the longer
experimental series. As for the level of dispersion, it is 4 percentage points
Actual: Michigan
Actual: Michigan Ex
o
wxa
o
9i i
\h
6-
? 3H
\1
v^ v
li Vv
iff"
o1950
1960
1970
1980
1990
2000
Year
higher on average in the data than predicted by the model. This may be
partially accounted for by some measurement error in the construction of
the Michigan series. More likely, however, it reflects idiosyncratic heterogeneity in the population that is not captured by the model. Individuals in
the public probably differ in their sources of information, in their sophistication in making forecasts, or even in their commitment to truthful reporting in a survey. None of these sources of individual-level variation are
captured by the sticky-information model, but they might cause the high
levels of disagreement observed in the data.10
Section 4 outlined several stylized facts regarding the dispersion of
inflation expectations in the survey data. The interquartile range of
expected inflation was found to rise with inflation and with the squared
change in annual inflation over the last year. The output gap did not seem
to affect significantly the dispersion of inflation expectations. We reestimate the regressions in panels A and C of Table 6, now using as the
10. An interesting illustration of this heterogeneity is provided by Bryan and Ventaku (2001),
who find that men and women in the Michigan Survey have statistically significant different expectations of inflation. Needless to say, the sticky-information model does not
incorporate gender heterogeneity.
Bivariate regressions
Constant
Inflation rate
AInflation-squared
Output gap
Adjusted R2
N
0.005***
(0.001)
0.127***
(0.028)
3.581***
(0.928)
0.009
(0.051)
0.469
579
0.166***
(0.027)
6.702***
(1.389)
0.018
(0.080)
579
I. *** denotes statistical significance at the 1% level (Newey-West standard errors in parentheses; correcting for autocorrelation up to one year).
Actual: Michigan
Actual: Livingston
9-
I 3H
o1950
1960
1980
1970
Year
1990
2000
0.262%
(0.310)
p: E M 2 [TCJ
0.436*
(0.261)
-1.416%*
(0.822)
0.088
No
p = 0.227
a: constant
Adj. R2
Reject efficiency?
a = (3 = 0
- 6 0 4 ***
(0.124)
0.107%
(0.211)
0.361
Constant
Adj. R2
+ 1 Jtf-13 +
a: constant
P: E M 2 [TCJ
7: inflation,
13
K: Treasury t>illf_13
5: unemployment,..^
Reject efficiency?
y=K = 8 = 0
Adjusted R2
h-U + 5 LJt-13
1.567%*
(0.824)
0.398
(0.329)
0.506***
(0.117)
-0.413**
(0.139)
-0.450***
(0.135)
Yes
p = 0.000
0.369
1. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively
(Newey-West standard errors in parentheses; correcting for autocorrelation up to one year).
Inflation
p(l): sum of 8 coefficients
Unemployment
y: date of forecast
K : 3 months prior
Treasury bill rate
8 : date of forecast
<)): 3 months prior
Reject adaptive expectations?
( Y = K = 8 = (|> = 0 )
Adjusted R2
N
1.182***
(0.100)
-0.561***
(0.087)
0.594***
(0.078)
0.117***
(0.026)
0.160***
(0.027)
Yes
p = 0.000
0.954
579
Bivariate
regressions
1.027***
(0.612; 1.508)
-0.009
(-0.078; 0.061)
0.029***
(0.004; 0.058)
-0.019
(-0.137; 0.108)
Reject at 5% level in
98.2% of histories
0.162
588
th
th
-0.010
(-0.089; 0.071)
0.030***
(0.005; 0.059)
-0.023
(-0.163; 0.116)
588
1. *** denotes statistical significance at the 1% level. (The 5 and 95 percentile coefficient estimates across
500 alternative histories are shown in parentheses.) Adjusted R2 refers to the average adjusted R2
obtained in the 500 different regressions.
0.057%
(-0.264; 0.369)
0.308**
(0.002; 0.6971)
-1.018%
(-2.879; 0.253)
Reject at 5% level in
95.4% of histories
Panel C: Are forecasting errors persistent? nt - Et_u 7if = a + (3 (rc M2 - (_24 7iM2)
P:n H2 -E f _ 24 [7t,_ 12 ]
a : constant
Adjusted R2
0.260***
(0.094; 0.396)
0.039%
(-0.237; 0.279)
0.072
a : constant
P : Ef-12 [nt]
Y: inflation,_13
K: Treasury billf_13
8 : output g a p w 3
Joint test on macro data (Y = K = 8 = 0)
Adjusted R2
N
-0.617%
(-3.090; 1.085)
0.032
(-0.884; 0.811)
0.064
(-0.178; 0.372)
0.068
(-0.185; 0.385)
0.170
(-0.105; 0.504)
Reject at 5% level in
78.6% of histories
0.070
569
1. *** and ** denote statistical significance at the 1% and 5% levels, respectively. (The 5th and 95th percentile
coefficient estimates across 500 alternative histories are shown in parentheses.) Adjusted R2 refers to the
average adjusted R2 obtained in the 500 different regressions.
1.100**
(0.177; 2.082)
0.380**
(0.064; 0.744)
-0.300
(-0.775; 0.190)
0.063
(-0.042; 0.165)
0.149
(-0.111; 0.371)
Reject at 5% level
in 100% of histories
0.896
569
1. ** denotes statistical significance at the 5% level. (The 5th and 95th percentile coefficient estimates
across 500 alternative histories are shown in parentheses.) Adjusted R2 refers to the average adjusted R2
obtained in the 500 different regressions.
rate of 11%, one of the highest in the postwar United States. Over the next
three years, using contractionary monetary policy, he sharply reduced the
inflation rate to 4%. This sudden change in policy and the resulting shock
to inflation provides an interesting natural experiment for the study of
inflation expectations. The evolution of the distribution of inflation
expectations between 1979 and 1982 in the Michigan Survey is plotted
in Figure 12.13 For each quarter there were on average 2,350 observations in the Michigan Survey, and the frequency distributions are estimated nonparametrically using a normal kernel-smoothing function.
Three features of the evolution of the distribution of inflation expectations stand out from Figure 12. First, expectations adjusted slowly to this
change in regime. The distribution of expectations shifts leftward only
gradually over time in the data. Second, in the process, dispersion
increases and the distribution flattens. Third, during the transition, the
distribution became approximately bimodal.
We now turn to asking whether the sticky-information model can
account for the evolution of the full distribution of expectations observed
in the survey data during this period. Figure 13 plots the distribution of
13. The Livingston and SPF surveys have too few observations at any given point in time to
generate meaningful frequency distributions.
1979, Q2
1979, Q3
1979, Q4
1980, Ql
1980, Q2
1980, Q3
1980, Q4
1981,Ql
1981, Q2
1981, Q3
1981, Q4
1982, Ql
1982, Q2
1982, Q3
1982, Q4
0.0080.0060.0040.0020.000 -
n of Pop
o
"S
o
o
iH
0.0080.0060.0040.002-
o.ooo0.0080.0060.004 J
0.002-
o.ooo0.0080.0060.0040.002-
o.ooo-5
10
15
20
-5
10
15
20
-5
10
15
20
-5
10
15
20
1979, Q2
1979, Q3
1979,
Q4
1980, Ql
1980, Q2
1980, Q3
1980,
Q4
1981, Q2
1981, Q3
1981,
Q4
1982, Q2
1982, Q3
1982,
Q4
0.60.40.2-
o.o0.6-
0.4-
0.2-
Q,
O
o.o-
i^
M
1981,Ql
i-i
oa
o
0.60.40.2-
IH
/I \
o.o-
1982, Ql
12
12
12
12
7. Conclusion
Regular attendees of the NBER Macroeconomics Annual conference are well
aware of one fact: people often disagree with one another. Indeed, disagreement about the state of the field and the most promising avenues for research
may be the conference's most reliable feature. Despite the prevalence of disagreement among conference participants, however, disagreement is conspicuously absent in the theories being discussed. In most standard
macroeconomic models, people share a common information set and form
expectations rationally. There is typically little room for people to disagree.
Our goal in this paper is to suggest that disagreement may be a key to
macroeconomic dynamics. We believe we have established three facts
about inflation expectations. First, not everyone has the same expectations.
The amount of disagreement is substantial. Second, the amount of disagreement varies over time together with other economic aggregates.
Third, the sticky-information model, according to which some people form
expectations based on outdated information, seems capable of explaining
many features of the observed evolution of both the central tendency and
the dispersion of inflation expectations over the past 50 years.
Prior to 1966, the survey did not probe quantitative expectations at all,
asking only the qualitative question.
Thus, for the full sample period, we have a continuous series of only
qualitative expectations. Even the exact coding of this question has varied
through time (Juster and Comment, 1978):
1948 (Ql)-1952 (Ql): "What do you think will happen to the prices of
the things you buy?"
1951 (Q4), 1952 (Q2)-1961 (Ql): "What do you expect prices of
household items and clothing will do during the next year or sostay
where they are, go up or go down?"
.75-
.5-
.25-
o1950
1960
1970
1980
Year
1990
2000
Thus, we have two independent data points for each month (%Same is
perfectly collinear with %Up+%Down), and we would like to recover two
time-varying parameters. The above two expressions can be solved simultaneously to yield:
F-l{%Downt) + FN\l-%UVt)
{%Downs)-F-l{l-%Upt)
Gt
= C
Not surprisingly, we can recover the time series of the mean and standard deviation of inflation expectations up to a multiplicative parameter,
c; that is, we can describe the time series of the mean and dispersion of
inflation expectations, but the scale is not directly interpretable. To
recover a more interpretable scaling, we can either make an ad hoc
assumption about the width of the zone from which same responses are
drawn, or fit some other feature of the data. We follow the second approach
and equate the sample mean of the experimental series and the corresponding quantitative estimates of median inflation expectations from the same
survey over the shorter 1978-2001 period when both quantitative and
qualitative data are available. (We denote the median inflation expectation by it.)14 formally, this can be stated:
1978-2001
1978-2001
/_, ^
zL71
c=
-i
1
(%
(%Down
t) + F" (%1 - Upt
FNl(%Downt)-F-\%l-UPt)
This assumption yields an estimate of c = 1.7%. That is, the specific scaling adopted yields the intuitively plausible estimate that those expecting
inflation between -1.7% and +1.7% respond that prices will stay where
they are now. More to the point, this specific scaling assumption is not
crucial to any of our regression estimates. It affects the interpretation of
the magnitude of coefficients but not the statistical significance.
Thus, for our sample of T periods, with 2T + 1 parameters and 2T + 1
unknowns, we can estimate the time series of the mean and standard
deviation of inflation expectations. As a final step, we rely on the assumption of normality to convert our estimate of the sample standard deviation
into an estimate of the interquartile range.
Figures 1 and 3 show that the median and interquartile range of the
constructed series move quite closely with the quantitative estimates over
the period from 1978. Table 2 reports on the correlation of this series with
other estimates.
REFERENCES
Ball, Laurence. (2000). Near-rationality and inflation in two monetary regimes.
Cambridge, MA: National Bureau of Economic Research. NBER Working Paper
7988.
Ball, Laurence, and Dean Croushore. (2003). Expectations and the effects of monetary policy. Journal of Money, Credit and Banking 35(4): 473-484.
Ball, Laurence, N. Gregory Mankiw, and Ricardo Reis. (2003). Monetary policy for
inattentive economies, Journal of Monetary Economics, forthcoming.
Barnea Amir, Amihud Dotan, and Josef Lakonishok. (1979). The effect of price
level uncertainty on the determination of nominal interest rates: Some empirical evidence. Southern Economic Journal 46(2): 609-614.
Bomberger, William, and William Frazer. (1981). Interest rates, uncertainty and the
Livingston data. Journal of Finance 36(3): 661-675.
14. It is just as valid to refer to the mean of this experimental series as the median expectation, given the assumption of normality.
Makin, John. (1982). Anticipated money, inflation uncertainty and real economic
activity. Review of Economics and Statistics 64(1):126-134.
Makin, John. (1983). Real interest, money surprises, anticipated inflation and fiscal deficits. Review of Economics and Statistics 65(3):374-384.
Mankiw, N. Gregory, and Ricardo Reis. (2002). Sticky information versus sticky
prices: A proposal to replace the new Keynesian Phillips curve. Quarterly Journal
of Economics 117(4):1295-1328.
Mankiw, N. Gregory, and Ricardo Reis. (2003). Sticky information: A model of
monetary non-neutrality and structural slumps. In Knowledge, Information and
Expectations in Modern Macroeconomics: In Honor of Edmund S. Phelps, P. Aghion,
248 KING
Comment
ROBERT G. KING*
Boston University, NBER, and Federal Reserve Bank of Richmond
Comment 249
250 KING
move over the course of business cycles; and how they evolved during an
important episode, the Volcker deflation. It is sure to stimulate much
interesting future research.
1.3 LINK TO THE STICKY-EXPECTATIONS MODEL
The sticky-expectations model of MR implies that macroeconomic
shocksparticularly monetary policy shockshave real effects because
some agents adjust expectations and others don't. It also has the effect
that monetary shocks cause dispersion in inflation expectations because
some agents adjust their forecasts immediately when a shock occurs and
others do so only gradually. This implication motivates the current paper:
MRW want to find out whether there are important changes over time in
the cross-sectional variability of expectations.
Now, there are other empirical implications of the sticky-expectations
model that one might want to explore, both at the micro and macro levels. In the MR model, when a firm gets an opportunity to update its information, it chooses an entire path for future nominal prices that it will
charge until its next information update.1 To me, this micro implication flies
in the face of one of the central facts that new Keynesian macroeconomics
has long stressed, which is the tendency for many nominal prices to stay
constant for substantial periods of time.2
And there are also macro implications of this adjustment pattern. Ball,
Mankiw, and Romer (1988) use data on a large number of different
countries to argue for sticky-price models rather than the alternative
information-confusion model of Lucas. They argue that sticky-price
models imply that the output-inflation trade-off should depend negatively on the average rate of inflation because high rates of inflation
would induce firms to undertake more frequent adjustments. They
argue that cross-country evidence strongly supports this implication,
rather than Lucas's implication that the slope should depend on variability of inflation. Now, because a Mankiw-Reis firm sets a path of
prices, it can neutralize the effects of the average inflation rate on its
real revenues. Its incentives for frequency of information adjustment
would therefore be unaffected by average inflation, just like Lucas's
flexible price firm, because the MR firm's price is flexible with respect
to forecasted inflation.
2. If a firm chose a nominal price that would be held fixed until the next receipt of information, then the MR model has the attractive characteristic that it simply collapses to the
well-known Calvo (1983) model. Thus, the essential feature of the model is that the firm
chooses a price plan rather than a price.
2. A notable and important exception is the practices of selective discounts, such as sales.
Comment 251
252 KING
1.6 CRITICAL QUESTION
As macroeconomists, we know that there are lots of types of heterogeneity in the world. We abstract from many in building our models, as Muth
did. The key to successful macro model building is to put in heterogeneity that is important for the issue at hand and to leave out the rest. For
example, in studying capital formation, one might think that it is important to take careful account of the age distribution of capital stocks
because this distribution could aid in predicting the timing of firms'
upgrades and replacements. One would like this age distribution of capital stocks to reflect underlying costs, presumably of a fixed sort, that keep
firms from rapidly adjusting their capital stocks.
More specifically, one might thinkas I didthat lumpy investment at the micro level would produce an important set of distributed
lag effects on aggregate investment not present in standard neoclassical models. But the general equilibrium analysis of Thomas (2002)
shows that this need not be the case: one carefully constructed model
with a rich age distribution of capital stocks does not produce very
different investment behavior than the simplest neoclassical model.
So this form of heterogeneity did not turn out to be important in one
particular setting.
By contrast, modern sticky-price models take heterogeneity in nominal
prices, a comparable age distribution of prices, as a first-order phenomenon. Many such models produce very different real responses from those
in flexible price models without a distribution of prices. While most of
these studies impose a time-dependent pattern of price adjustment, the
nonneutrality results of some sticky-price models survive the introduction of state dependent pricing.
So the critical question becomes, Is heterogeneity in beliefs important
for macroeconomic models of the Phillips curve?
Comment 253
+1
=atnt
1+(l-at)n
3. A coefficient (3 is sometimes inserted before expected future inflation. Because it is a quarterly discount factor, its value is just below 1 and it is omitted for simplicity in the discussion below.
254 KING
This expression captures two effects familiar from Ball's work on sticky
prices and deflation. First, a perfectly credible (at = 1) deflation produces
a boom in output. Second, if inflation is reduced in an imperfectly credible
manner, then a recession will occur. For example, if the disinflation is
completely incredible (at = 0), then output declines in lock-step with inflation. It also highlights that the behavior of output depends importantly on
the dynamics of beliefs that are here taken to be common across all agents,
i.e., on the behavior of the at over time.
2.2 DYNAMICS OF A SUCCESSFUL DEFLATION
I now use this simple model to analyze an example of dynamics within a
successful disinflation, which is assumed to take three years to complete
and to reduce the inflation rate from 10% per year to 4% per year. In particular, I suppose that the beliefs about the disinflation are initially stubborn, with at = 0 for three quarters, and then gradually rise until the
disinflation is fully credible at its endpoint.
The particular assumptions and their implications for output are displayed in Figure 1. Inflation is assumed to decline gradually, as displayed
in the top panel. The credibility of the disinflation rises through time.
Figure 1 DYNAMICS OF AN IMPERFECTLY CREDIBLE BUT ULTIMATELY
SUCCESSFUL DEFLATION
10
Comment 255
Output is initially not much affected but then declines because the disinflation is incredible.4 As its credibility rises, the disinflation's real consequences evaporate.
2.3 REINTERPRETING THE MODEL
To this point, we have assumed that there is no disagreement about inflation expectations in the sense of Mankiw, Reis, and Wolfers. Suppose,
however, that we now assume that there is such disagreement. A fraction
at of the population is optimistic, believing that the disinflation will continue, while the remaining population members are pessimistic. Under
this alternative interpretation, the dynamics of inflation and output are
unaffected, but there would be variability in measures of disagreement
similar to those considered in this paper: disagreement would be small at
the beginning and end of the disinflation, while it would be higher in the
middle. So, in this model, disagreement about inflation expectations can
occur and evolve over time. But modeling these disagreements does not
seem essential to understanding the episode.
2.4 CONNECTING WITH THE ACTUAL DISINFLATION EXPERIENCE
I think that the actual disinflation experience involved the following four
features:
First, it was widely discussed: it is hard to imagine that agents didn't know
that something was up. 5
Second, it was widely debated on two dimensions. People disagreed about
whether it would work and whether it was a good idea. The former
suggests disagreement about expectations.
Third, there was some uncertainty about what was going on: people were
not sure what the Federal Reserve was up to in terms of its long-range
objectives for inflation.
Fourth, it was imperfectly credible. As Shapiro (1994) notes, the Volcker
deflation is very different from the prior disinflation attempts by the
4. As the reader will note, the scale of the output effect depends entirely on the choice of the
parameter <j). In drawing the graph, I chose a = .2, which meant that a maximum output
decline of about 2.5% occurred, although no vertical scale is included in the diagram. A
choice of (|) = .05 would have alternatively brought about a maximum 10% decline in output. In models that derive from underlying micro structure, it is related to two deeper
parameters: the effect of real marginal cost on inflation (which Gali and Gertler [1999] estimate to be about .05) and the elasticity of real marginal cost to the output gap. Dotsey and
King (2001) discuss how some structural features of the underlying economy affect this
latter feature. Values of this elasticity much less than 1 arise from models with elastic labor
supply, variable capacity utilization, and intermediate inputs. Hence, small values of <j)
and large output effects are not hard to generate from modern sticky-price models.
5. This restates a common criticism of the Barro-Lucas-type incomplete information models,
which my then-colleague Stan Engerman once summarized as "Don't the people in your
economies have telephones?" and Ed Prescott later put as "People read newspapers."
256 KING
Federal Reserve. Within a few years after each of the four prior
episodes, inflation was reduced only temporarily and then returned to
an even higher level within a few years.
During the Volcker deflation, long-term interest rates stayed high for a
long time, much longer than any modern pricing modelincluding that
of Mankiw and Reiswould predict, if there was not imperfect credibility about long-term inflation. Unraveling the nature of this episode is an
important topic for research in monetary economics, but I am not convinced that understanding the dynamics of measures of disagreement
about expectations is important for understanding the episode.
2.5 IMPERFECT CREDIBILITY VERSUS STICKY EXPECTATIONS
In terms of practical macroeconomics, one might ask whether I have drawn
a distinction without a difference in my discussion. I don't think so. Sticky
expectations are a structural feature of price dynamics for Mankiw and Reis
and describe both normal situations and unusual events. Imperfect credibility is a feature of the macroeconomy and monetary policy, and is likely
to be more important in some situations than others. So, the inflation-output trade-off during the Volcker deflation might give a poor guide to the
nature of that trade-off in a current monetary policy context.
REFERENCES
Ball, Laurence. (1994). Credible disinflation with staggered price-setting. American
Economic Review 84(1): 282-289.
Ball, Laurence. (1995). Disinflation with imperfect credibility. Journal of Monetary
Economics 35(1): 5-23.
Ball, Laurence, N. Gregory Mankiw, and David Romer. (1988). The new Keynesian
economics and the output-inflation trade-off. Brookings Papers on Economic
Activity 1988(1): 1-82.
Barro, Robert J. (1976). Rational expectations and the role of monetary policy.
Journal of Monetary Economics 2(1): 1-32.
Dotsey, Michael, and Robert G. King. (2001). Production, pricing and persistence.
Cambridge, MA: National Bureau of Economic Research. NBER Working Paper
No. 8407.
Gali, Jordi, and Mark Gertler. (1999). Inflation dynamics: A structural econometric
analysis. Journal of Monetary Economics 44(2): 195-222.
Gordon, Robert J. (1970). The recent acceleration of inflation and its lessons for the
future. Brookings Papers on Macroeconomics 1:8-41.
Lucas, Robert E., Jr. (1972). Econometric testing of the natural rate hypothesis. In
The Econometrics of Price Determination, Otto Eckstein (ed.). Washington, DC:
Comment 257
Mankiw, N. Gregory, and Ricardo Reis. (2002). Sticky information versus sticky
prices: A proposal to replace the new Keynesian Phillips curve. Quarterly Journal
of Economics 117(4): 1295-1328.
Muth, John R (1961). Rational expectations and the theory of price movements.
Econometrica 29(3):315-335.
Sargent, Thomas J. (1971). A note on the accelerationist controversy. Journal of
Money, Credit and Banking 3(August):50-60.
Shapiro, Matthew D. (1994). Federal Reserve policy: Cause and effect. In Monetary
Policy, N. G. Mankiw (ed.). Cambridge, MA: MIT Press.
Solow, Robert M. (1969). Price Expectations and the Behavior of the Price Level.
Comment
JOHN C. WILLIAMS
Federal Reserve Bank of San Francisco
258 WILLIAMS
Gurkaynak et al. (2003) find that forward nominal interest rates are
highly sensitive to economic news, and they provide evidence that this
sensitivity is primarily related to the inflation component of interest
rates. Because financial market participants "put their money where
there mouths are," one is tempted to put greater faith in estimates taken
from financial market data, even while recognizing the difficult measurement problems associated with extracting expectations from these
data. Still, additional study and comparison of both sources of expectations data is needed to form a more complete picture of the properties of
expectations.
The remainder of my discussion will focus on two topics: learning and
model uncertainty. The first relates to is the real-time information that
forecasters are assumed to possess. The second provides an alternative
explanation of the evidence on dispersion in forecasters' inflation expectations based on the notion that there exists a range of competing forecast
models. I find that model uncertainty provides an intuitively more
appealing description of the form of disagreement among economists
than that proposed in the paper.
The authors argue that the Mankiw and Reis (2002) sticky-information
model can explain many of the properties of the median values and dispersion of surveys of inflation expectations. In this model, agents use a
three-variable, 12-lag monthly vector auto regression (VAR), that includes
inflation, the output gap, and the three-month T-bill rate, estimated over
the entire postwar sample to generate forecasts.1 Individual agents, however, update their expectations only at random intervals, with a 10% probability of an update in each month. Given this structure, the resulting
median sticky-information forecast is closely related to the median of a
geometrically weighted average of past inflation forecasts. The crosssectional dispersion in forecasts reflects the dispersion of forecasts across
vintages. In fact, there is absolutely no disagreement about forecasting
methods; all the differences arise from the differences across vintages of
forecasts that people are assumed to use.
I find the assumption that households and economists had access to
the full-sample VAR estimates to be unrealistic: people in 1960 simply
did not possess the knowledge of forecast model specification and
1. The use of the output gap in the forecasting VAR is problematic because of well-documented problems with real-time estimates of GDP and potential (or the natural rate of)
output, issues emphasized by Orphanides (2001), Orphanides and van Norden (2002),
and others. A preferable approach would be to use the unemployment rate or the realtime estimates of capacity utilization, which is the approach I follow in the model-based
exercises reported in this discussion. As noted by the authors, their results are not sensitive to the use of the output gap in this application, so this criticism is intended more as
a general warning.
Comment 259
Figure 1 IN SAMPLE VERSUS OUT-OF-SAMPLE FORECASTS
18Livingston Survey
VAR (Full-Sample)
VAR (Real-Time)
1614121086420-
1960
1965
1970
1975
1980
1985
1990
1995
2000
Year
260 WILLIAMS
the current quarter.3 The dashed black line reports the corresponding forecasts from a VAR estimated over the full sample. For comparison, the
solid black line shows the Livingston Survey of expected price increases
over the next 12 months.
The forecasts of the real-time and full-sample VARs are nearly identical
from the mid-1980s on, but in the earlier period they display sizable differences. In the 1960s and 1970s, the real-time VAR tracks the Livingston
Survey closely, while the full-sample VAR significantly overpredicts inflation expectations nearly throughout the period. In contrast, in the early
1980s, during the Volcker disinflation, the real-time VAR severely overpredicts inflation expectations. This discrepancy may reflect judgmental
modifications to forecasts that incorporate extra-model knowledge of the
Fed's goals and actions at the time (as well as other influences on inflation) not captured by the simple VAR.
The issue of real-time forecasts versus forecasts after the fact also has
implications for the interpretation of forecast rationality tests reported
in the paper. In describing the properties of median forecasts, the
authors apply standard tests of forecast rationality to the four survey
series that they study. Such tests boil down to looking for correlations
between forecast errors and observable variables, the existence of which
implies that forecast errors are predictable and therefore not rational.
They consider four such tests. The simplest test is that for forecast bias,
i.e., nonzero mean in forecast errors. A second test is for serial correlation in forecast errors in nonoverlapping periods. A third test, which I
call the forecast information test in the following, is a test of correlation
between forecast errors and a constant and the forecast itself. The final
test, which I call the all information test, is a joint test of the correlation
between forecast errors and a set of variables taken from the VAR
described above, assumed to be in forecasters' information set.
They find mixed results on bias and forecast information tests, but
rationality of the median value of surveys is rejected based on the serial
correlation and all information tests. They then show that the median
forecast predicted by the sticky-information model yields similar
resultswith forecast errors exhibiting positive serial correlation and a
high rate of rejection of the forecast information and all information
testsproviding support for that model.
An alternative interpretation of these results is that forecasters have been
learning about quantitative macroeconomic relationships over time. The
3. I chose the CPI for this analysis because it sidesteps the issue of differences between real-time
and final revised data in national income account price indexes, such as the GDP deflator.
The CPI is not revised except for seasonal factors, and because I am focusing on four-quarter
changes in prices, seasonal factors should be of little importance to the analysis.
Comment 261
tests are based on the correlations in the full sample and ignore the fact that
forecasters, even those with the correct VAR model, had inaccurate estimates of these relationships at the time of their forecasts. Because of sampling errors in the forecaster's model, these tests are biased toward rejecting
the null of rationality.4
Indeed, a wide variety of reasonable forecasting models, including a quarterly version of the VAR used in the paper (with the unemployment rate
substituting for the output gap), yield a pattern of rejections similar to those
seen in the survey data when one assumes that the forecasts were constructed in real-time using knowledge of the data correlations available at
the time. Table 1 reports the results from rationality tests from several simple forecasting models. The first line of the table reports the results from the
three-variable VAR with four quarterly lags, where the VAR is reestimated
each period to incorporate the latest observed data point. This real-time VAR
exhibits no bias over the past 40 years, but forecast rationality is rejected
based on positive forecast error serial correlation and the two information
tests. (For this test, I include the most recent observed value of the inflation
rate, the unemployment rate, and the 3-month T-bill rate.) The results for
other forecast models (the details of which I describe below) are also consistent with the evidence from the surveys. And as indicated in the bottom line
of the table, the median forecast among these 10 forecasting models also
exhibits the pattern of rejections seen in the survey data. This evidence suggests that forecasters use models in which the parameters change over time.
I now turn to the second half of the paper. The authors show that the
sticky-information model provides a parsimonious theory of disagreement that is in accord, at least qualitatively, with the time-series pattern of
disagreement seen in the survey data. In addition, the model can generate
a positive correlation between disagreement and the inflation rate, also a
prominent feature of the survey data.
The evidence for the sticky-information model from inflation expectations disagreement from household surveys, however, is not clear cut. The
model cannot come close to matching the magnitude of the dispersion in
household inflation expectations, for which the interquartile range
(IQR)that already excludes one half of the sample as outlierscan
reach 10 percentage points! In Figure 10 of the paper, the difference
between the measure of disagreement in the data and that predicted by
4. In addition to uncertainty about model parameters, the specification of forecasting models changes over time in response to incoming data, driving another wedge between the
information set available to real-time forecasters and after-the-fact calculations of what
forecasters "should have known." In models with time-varying latent variables such as
the natural rates of unemployment and interest, this problem also extends to the real-time
specification and estimation of the latent variable data-generating processes, as discussed
in Orphanides and Williams (2002).
erro
e, th
isth
Ol
low
QJ
IT
u?
.2
:en
stei
en 0)
to X i
en
to
u
QJ
Xj Xi
"*^
T3
tO
"c
u
to
E denotes roo
ata.
asteris ks, between'.
ion
ecast erro:
res
forecast error
sta ndard err<
CN
CN
O
^H
>
5Xi
T3
to
oopooooooo
tO
CJ
01 QJ
-a
c
QJ
to 6 0
QJ 6 0
01 to
g 1O
S K
ean squaj
percent;
onstant a
a consta
_3
to "to
>
in
01
"o
-a o
j>-
oooooooooo
III
I
CN
O
en
'35 In
60
uart
cent
inr<
TH
0)
60
o^
tO
oi O -2 ^
00 ^O ON
CN O O
CN CN CN
-2 o3 -a
6OT3 X!
60 QJ to
<
J i m o Jiifio
Opp
Opp
Opp
ZocsZooZoo
S H CO C T H CO
J J L1 O1O
S " ^
w
CD
<
pa
Sw
OJ
o
T3
O fS
Comment 263
the model is not constant over time and appears to be highly persistent.
There's clearly something else going on here, with the sticky-information
model capturing only part of the process of households' expectations
formation.
The sticky-information model is closely linked to Chris Carroll's (2003)
model, whereby households randomly come into contact with professional forecasts. I find his model to be a highly plausible description of
expectations formation by households, who are unlikely to keep in constant touch with the latest macroeconomic data. But as a macroeconomic
forecaster myself, I find it entirely implausible as a description of the
behavior of business economists and professional forecasters surveyed in
the Livingston Survey and the Survey of Professional Forecasters (SPF),
respectively. Professional forecasters update their forecasts regularly and
update their forecasting models at frequent intervals. The primary reason
economists' forecasts disagree is not due to lags in formulating new forecasts, but instead is because economists themselves disagree about how to
model the economy best!5 This is an aspect of dispersions in expectations
entirely absent from the model in the paper.
In fact, there already exist theories of rational heterogeneity of beliefs
that naturally yield expectations disagreement (see, for example, Brock
and Hommes [1997], Branch [2003], and Branch and Evans [2003]). These
theories assume that agents have at their disposal a range of forecasting
models but are uncertain about which model or models to use. They
update their model choice or priors over the various models based on
forecasting performance. Idiosyncratic differences in agents' characteristics, say, different initial conditions in model priors and the costs for learning new models, implies that a range of models will be in use at any point
in time.
This description matches closely the real-world practice of economic
forecasting that recognizes the high degree of model uncertainty in forecasting. There exists many competing inflation forecast models, including
time-series models, Bayesian VARs, reduced-form Phillips curves, and
large-scale macroeconometric models in use at the same time. And for
each model, several variants differ with respect to model specification and
estimation, including the lag length of explanatory variables; treatment of
latent variables; sample size; and the inclusion or exclusion of additional
explanatory variables such as energy prices, import prices, price control
dummies, sample, wages, productivity, etc. (compare Brayton et al. [1999]
and Stock and Watson [1999]). These models have similarly good track
5. The disagreement seen in published forecasts is an imperfect measure of true disagreement. There are incentives both not stray too far from the consensus (Scharfstein and
Stein, 1990; Lamont, 2002) as well as to stand out from the crowd (Laster et al., 1999).
264 WILLIAMS
records in terms of forecasting performance, but at times they can yield
strikingly different forecasts. In practice, forecasters combine the forecasts
from a subset of these models, along with extra-model information, to
arrive at a point estimate forecast.
But why don't all forecasters arrive at the same forecast? For the same
reason as in the theory sketched above: idiosyncratic differences between
economists imply persistent deviations in modeling choices. One source
of such differences might originate during graduate school training. For
example, economists trained in the 1960s likely rely more heavily on
structural macroeconometric models for forecasting, while those trained
during the 1990s probably place more weight on Bayesian VARs. Because
these models are about equally good in terms of forecasting accuracy, the
pace of convergence to a single best mix of models is likely to be slow.
To illustrate how model uncertainty can lead to forecast disagreement
conforming to the evidence presented in the paper, I construct an artificial
universe of forecasters, each of whom is assumed to use one of the 10 forecasting models listed in Table 1. As seen in the table, these models have
roughly similar out-of-sample forecast accuracy. In each case, the model is
reestimated each period. The parameters are unrestricted. For each of the
three main types of models, I consider three variants: one is estimated by
standard ordinary least squares (OLS) and the second and third are estimated by weighted least squares (WLS), with the weights declining geometrically using the values 0.985 or 0.970, as indicated in the table. WLS
estimation is designed as protection against structural change by downweighting old data (Evans and Honkapohja, 2001). The first main model
is the VAR described above. The second model is a Phillips curve model
that includes a constant, four lags of inflation, and two lags of the unemployment rate. The third model is a fourth-order autoregression. The set
of models is completed with a simple random walk model where the forecast inflation over the next four quarters equals the inflation rate over the
past four quarters.
The out-of-sample forecasting performance (measured by the root
mean squared forecast error) of these various models is quite similar. The
random walk model beats the other models without discounting, consistent with the findings of Atkeson and Ohanion (2001). The VAR model is
the worst performer of the group, supporting Fair's (1979) finding that
unrestricted VAR models tend to perform poorly out of sample. The
Phillips curve model, with discounting, is the best performer of the group,
and in all three cases the model with discounting outperforms the OLS
version. Evidently, structural breaks are of great enough importance in
this sample that protecting against such breaks outweighs the efficiency
loss associated with discarding data (see Orphanides and Williams
Comment 265
[2004]). Finally, the median forecast from this set of 10 models performs
better than any individual forecast, in line with the literature that averaging across forecast models improves performance (see Clement [1989] and
Granger [1989] for surveys of this topic).
At times, these forecasting models yield different forecasts of inflation,
as shown by the shaded region in Figure 2. The degree of disagreement
across models widens appreciably around 1970, again in the mid-1970s,
and most strikingly during the period of the disinflation commencing at
the end of the 1970s and continuing into the early 1980s. The magnitude
of disagreement across models is much smaller during the 1960s and from
the mid-1980s through the end of the sample.
The time-series pattern seen in the interquartile range from these forecast models is similar to that seen in the Livingston Survey IQR, as
depicted in Figure 3. During periods of stable and low inflation, forecast
dispersion among these models is modest, but it spikes when inflation is
high or changing rapidly, exhibiting the same pattern as in the IQR from
the survey data. Of course, the purpose of this exercise is only to illustrate
how model uncertainty can give rise to forecast disagreement similar to
that seen in the survey data. As noted above, the extent of model disagreement extends beyond the set of models I have considered here,
Figure 2 MODEL UNCERTAINTY AND FORECAST DISAGREEMENT
20 -i
Livingston Survey
Range of Model Forecasts
Midpoint of Range
Sticky-Information Model
15-
10-
5-
0-
-5
I960
1965
1970
1975
1980
Year
1985
1990
1995
2000
266 WILLIAMS
1960
1965
1970
1975
1980
Year
1985
1990
1995
2000
Comment 267
268 WILLIAMS
Orphanides, Athanasios, and John C. Williams. (2004). Imperfect knowledge,
inflation expectations, and monetary policy. In Inflation Targeting, Michael
Woodford, (ed.). Chicago, IL: University of Chicago Press.
Scharfstein, David S., and Jeremy C. Stein. (1990). Herd behavior and investment.
American Economic Review 80(3):465^479.
Stock, James H., and Mark W. Watson. (1999). Forecasting inflation. Journal of
Monetary Economics 44(2):293-335.
Discussion
A number of participants suggested alternative explanations for the disagreement across agents in inflation expectations documented by the
authors, and they recommended that the authors test their theory
against plausible alternatives. Olivier Blanchard suggested the possibility that people form inflation expectations based on small samples of
goods. Inflation variability across goods could then drive dispersion in
expectations. He noted that there is indeed a lot of inflation variability
across goods. He pointed out that this story is consistent with the fact
that there is more dispersion in inflation expectations when inflation is
changing. He also speculated that women might buy different baskets of
goods from men, hence explaining the difference in expected inflation
between women and men in the Michigan Survey. Justin Wolfers
explained that, in favor of Blanchard's story, people's assessment of past
inflation is a good predictor of their inflation expectation. The caveat is
that the extent of variation in the inflation that they think they have
experienced is not rationalizable. Ken Rogoff pointed out that there is
evidence of substantial inflation dispersion across cities within the
United States. He suggested that this might favor the hypotheses of
Blanchard and Williams: that individual forecasts might be made on the
basis of different baskets.
Another explanation was suggested by Mike Woodford. He remarked
that Chris Sims's theory of limited information-processing capacity could
endogenously generate the co-movement of disagreement and other
macro variables. Greg Mankiw responded that the sticky-information
model is a close relation of Chris Sims's work, but that it is analytically
more tractable and generates testable predictions more easily than the
work of Sims.
Ken Rogoff asked whether a story about policy credibility could explain
the fact that there is more disagreement when inflation is changing. He
hypothesized that fat tails in the distribution could be due to expectations
of more extreme histories than the United States has experienced. He
noted that, in view of the historically higher inflation experienced by both
Discussion 269
270 DISCUSSION
literature that documented incentives for forecasters to make extreme
predictions to attract attention. Greg Mankiw responded that the fact that
the cyclical response of expected inflation is as predicted by the model,
and that disagreement co-moves strongly across different surveys suggests that they are picking up more than just noise. Mankiw agreed that
the fact that private-sector forecasters are selling a product does affect the
incentives they face compared with, for example, the Fed's Green Book
forecast. Ken Rogoff noted that forecasters may try to avoid changing
their predictions to maintain credibility.
Ricardo Reis stressed that the main point of the paper was to demonstrate the extent of disagreement across agents in inflation expectations.
He remarked that nothing in macro theory so far can explain why there
should be so much disagreement, and the paper explored a first possible
explanation. Reis said that there is a large middle ground between adaptive and hyperrational expectations, and that the paper is an attempt to
explore one possible alternative. Greg Mankiw concluded that the bottom
line of the paper is that disagreement is widespread, and it varies over
time with variables that interest macroeconomists. Hence, macro models
should be consistent with disagreement of this type.